Detecting slender objects with uncertainty based on keypoint-displacement representation.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Standard object detection networks struggle with elongated, thin objects due to orientation challenges and bounding box limitations.
• Accurate detection of slender objects is crucial for quality control in manufacturing processes.

Purpose of the Study:

• To develop an object detection method specifically designed for slender objects.
• To improve the reliability and accuracy of slender object detection by incorporating uncertainty estimation.

Main Methods:

• A novel network architecture predicting keypoint heatmaps, displacement vector fields, and displacement uncertainty heatmaps.
• Representation of slender objects using a keypoint-displacement pattern, overcoming limitations of axis-aligned bounding boxes.
• Integration of an uncertainty branch to quantify detection reliability and refine training with ambiguous samples.

Main Results:

• Successful detection of slender objects, such as electrode sheet edges and electronic chip pins, in practical applications.
• Demonstrated ability to evaluate manufacturing quality based on detection results, including keypoint count, displacement, and uncertainty.
• The uncertainty branch enhances detection accuracy by down-weighting ambiguous training data.

Conclusions:

• The proposed keypoint-displacement pattern method effectively addresses challenges in detecting slender objects.
• Incorporating uncertainty estimation improves detection robustness and provides a reliable criterion for assessing results.
• The method shows significant potential for industrial quality inspection and manufacturing process optimization.